Many advanced computing systems, including supercomputers for example, utilize multiple computational units to improve performance in what is called a parallel system. The system of interconnections among parallel computational units is an important characteristic for determining performance. One technique for interconnecting parallel computational units involves construction of a communication network similar to a telephone network in which groups of network elements are connected to switching systems. The switching systems are interconnected in a hierarchical manner so that any switching station manages a workable number of connections.
One disadvantage of a network connection is an increase in the latency of access to another computational unit since transmission of a message traverses several stages of a network. Typically, periods of peak activity occur in which the network is saturated with numerous messages so that many messages simultaneously contend for the use of a switching station. Various network types have been devised with goals of reducing congestion, improving transmission speed and achieving a reasonable cost. These goals are typically attained by rapidly communicating between nodes and minimizing the number of interconnections that a node must support.
One conventional interconnection scheme is a ring of nodes with each node connected to two other nodes so that the line of interconnections forms a circle. The definition of a ring, in accordance with a standard definition of a ring network in the art of computing (IBM Dictionary of Computing, McDaniel G. ed., McGraw-Hill, Inc., 1994, p. 584) is a network configuration in which devices are connected by unidirectional transmission links to form a closed path. Another simple conventional scheme is a mesh in which each node is connected to its four nearest neighbors. The ring and mesh techniques advantageously limit the number of interconnections supported by a node. Unfortunately, the ring and mesh networks typically are plagued by lengthy delays in message communication since the number of nodes traversed in sending a message from one node to another may be quite large. These lengthy delays commonly cause a computational unit to remain idle awaiting a message in transit to the unit.
The earliest networks, generally beginning with telephone networks, utilize circuit switching in which each message is routed through the network along a dedicated path that is reserved for the duration of the communication analogous to a direct connection via a single circuit between the communicating parties. Circuit switching disadvantageously requires a lengthy setup time. Such delays are intolerable during the short and quick exchanges that take place between different computational units. Furthermore, a dedicated pathway is very wasteful of system bandwidth. One technique for solving the problems arising using circuit switching is called packet switching in which messages sent from one computational unit to another does not travel in a continuous stream to a dedicated circuit. Instead, each computational unit is connected to a node that subdivides messages into a sequence of data packets. A message contains an arbitrary sequence of binary digits that are preceded by addressing information. The length of the entire message is limited to a defined maximum length. A "header" containing at least the destination address and a sequence number is attached to each packet, and the packets are sent across the network. Addresses are read and packets are delivered within a fraction of a second. No circuit setup delay is imposed because no circuit is set up. System bandwidth is not wasted since there is no individual connection between two computational units. However, a small portion of the communication capacity is used for routing information, headers and other control information. When communication advances in isolated, short bursts, packet switching more efficiently utilizes network capacity. Because no transmission capacity is specifically reserved for an individual computational unit, time gaps between packets are filled with packets from other users. Packet switching implements a type of distributed multiplexing system by enabling all users to share lines on the network continuously.
Advances in technology result in improvement in computer system performance. However, the manner in which these technological advances are implemented will greatly determine the extent of improvement in performance. For example, performance improvements arising from completely optical computing strongly depend on an interconnection scheme that best exploits the advantages of optical technology.